Open Dreamer reproduces Dreamer 4 and publishes the full recipe
A JAX/Flax rebuild of Dreamer 4 you can train on one GPU, with a browser demo that plays a dreamed Minecraft — plus OpenSpace, an open skill hub for agents.
Copy markdown
A Dreamer 4 world model you can actually train
Reactor's Open Dreamer reimplements DeepMind's Dreamer 4 in JAX/Flax: a 1.6B-parameter, action-conditioned dynamics model plus a masked-autoencoder tokenizer that compresses each frame ~100× to 512 latents. Training pipeline, configs, and a local inference harness are all public — no weights gate, no waitlist.
Play the dream in your browser
A hosted demo streams a Minecraft world generated entirely by the model, with a Game⟷Dream toggle to flip between the real engine and the neural rollout. The action space is VPT-style — 27 binary buttons plus 121 mouse classes — so the weekend project is a controllable, generated game world you drive yourself.
The real gift is the stability recipe
The team says stability, not scale, ate most of their time, and they wrote the fixes down: Muon instead of LaProp, EMA weights for diffusion inference, float32 params with BF16 compute. CoinRun trains on a single GPU; Minecraft scales to a B200 at ~58% model-FLOPs utilization and ~24 GiB of model state.
OpenSpace: agents that get cheaper as they learn
Separately, HKUDS open-sourced OpenSpace, a skill hub that lets agents save, share, and reuse skills over MCP with lineage tracking. Maintainers report roughly 46% lower cost from reuse, and the repo has already cleared 4.7k stars.